@DynamicWebPaige whew... can scratch off "teach myself how to optimize GPU kernels" from my bucket list.
@dpetrou this is the dream! 😄
👀 ICYMI: K3 autonomously optimized GPU kernels in a 15-hour run, more than halving compute time, and built a compact Triton-like GPU compiler from scratch. Models are literally writing kernel code unsupervised now.
Users are excited about AI System K3 autonomously halving GPU kernel compute and training times because it delivers impressive optimizations that fulfill dreams and remove the need for tedious manual work.
Based on 3 visible X reactions from 6 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@DynamicWebPaige whew... can scratch off "teach myself how to optimize GPU kernels" from my bucket list.
@dpetrou this is the dream! 😄
👀 ICYMI: K3 autonomously optimized GPU kernels in a 15-hour run, more than halving compute time, and built a compact Triton-like GPU compiler from scratch. Models are literally writing kernel code unsupervised now.
Makes Anthropic’s decision to make Fable refuse your CUDA optimization requests look silly.
The system also built a custom GPU compiler from scratch.
Makes Anthropic’s decision to make Fable refuse your CUDA optimization requests look silly.
👀 ICYMI: K3 autonomously optimized GPU kernels in a 15-hour run, more than halving compute time, and built a compact Triton-like GPU compiler from scratch. Models are literally writing kernel code unsupervised now.
Self-evolving: AttnRes Kernel Optimization Given FLA Triton AttnRes at production scale (96 layers, 8192-dim model, 8192 tokens), the goal was to maximize training-side speed without changing numerics. Over 15 hours of nonstop iteration, K3 designed a novel two-phase kernel algorithm, fused kernels while preserving numerics, and reduced forward+backward time from 283.6 ms to 114.4 ms. K3 and Fable-5 (with potential fallback) reached similar performance, but K3 improved faster per iteration.
Makes Anthropic’s decision to make Fable refuse your CUDA optimization requests look silly.
Self-evolving: AttnRes Kernel Optimization Given FLA Triton AttnRes at production scale (96 layers, 8192-dim model, 8192 tokens), the goal was to maximize training-side speed without changing numerics. Over 15 hours of nonstop iteration, K3 designed a novel two-phase kernel algorithm, fused kernels while preserving numerics, and reduced forward+backward time from 283.6 ms to 114.4 ms. K3 and Fable-5 (with potential fallback) reached similar performance, but K3 improved faster per iteration.
Users are excited about AI System K3 autonomously halving GPU kernel compute and training times because it delivers impressive optimizations that fulfill dreams and remove the need for tedious manual work.
Based on 3 visible X reactions from 6 accounts; directional sample.
Ask a question below.
Published answers will appear here.
